added per-speaker samplers

.gitignore

@@ -7,4 +7,6 @@ __pycache__
 /*.egg-info
 /vall_e/version.py
 /build
-/.cache
+/.cache
+
+/vall_e/ext/interleaver.py

vall_e/config.py

@@ -162,6 +162,7 @@ class Model:
 	tasks: int = 8 # ["tts", "ns", "sr", "tse", "cse", "nse"] and leaves two more for anything else I want (like "svc")
 	arch_type: str = "transformer"
 	training: bool = True
+	interleave_pattern: str | None = None
 
 	@property
 	def full_name(self):

vall_e/data.py

@@ -12,6 +12,7 @@ import itertools
 
 from .config import cfg
 from .emb.qnt import trim, trim_random, repeat_extend_audio, merge_audio, decode_to_file
+from .utils.sampler import Sampler
 
 from collections import defaultdict
 from functools import cache, cached_property
@@ -173,6 +174,8 @@ class Dataset(_Dataset):
 		self.paths_by_spkr_name = _load_paths(self.dataset, self.dataset_type)
 		self.paths = list(itertools.chain.from_iterable(self.paths_by_spkr_name.values()))
 
+		self.samplers = { name: Sampler( paths, keep_all=True ) for name, paths in self.paths_by_spkr_name.items() }
+
 		if cfg.dataset.sample_type == "path":
 			self.paths = [*_interleaved_reorder(self.paths, self.get_speaker)]
 
@@ -215,6 +218,22 @@ class Dataset(_Dataset):
 	def tasks(self):
 		return cfg.dataset.tasks_list # ["tts", "tts", "ns", "sr", "tse", "tts", "tts"] # , "cse", "nse"
 
+	def save_state_dict(self, path):
+		state_dict = {
+			"samplers": { name: sampler.current_pool for name, sampler in self.samplers.items() }
+		}
+		torch.save(state_dict, path)
+
+	def load_state_dict(self, path):
+		state_dict = torch.load(path)
+
+		if "samplers" in state_dict:
+			# better than naively setting the entire object
+			for name, sampler in state_dict["samplers"].items():
+				if name not in self.samplers:
+					continue
+				self.samplers[name].current_pool = sampler
+
 	def _get_phone_symmap(self):
 		return get_phone_symmap()
 
@@ -290,7 +309,7 @@ class Dataset(_Dataset):
 		if cfg.dataset.sample_type == "speaker":
 			spkr_name = self.spkrs[index]
 			spkr_id = self.spkr_symmap[spkr_name]
-			path = random.choice([*set(self.paths_by_spkr_name[spkr_name])])
+			path = self.samplers[spkr_name].sample()
 		else:
 			path = self.paths[index]
 			spkr_name = self.get_speaker(path)
@@ -543,6 +562,10 @@ def create_datasets():
 	train_dataset = Dataset( training=True )
 	val_dataset = Dataset( phone_symmap=train_dataset.phone_symmap, training=False )
 
+	train_state_path = cfg.relpath / "train_dataset.pt"
+	if train_state_path.exists():
+		train_dataset.load_state_dict( train_state_path )
+
 	return train_dataset, val_dataset
 
 
@@ -752,6 +775,8 @@ if __name__ == "__main__":
 				del v[i]['resps']
 			print(f'{k}:', v)
 
+		train_dl.dataset.save_state_dict(cfg.relpath / "train_dataset.pt")
+
 	elif args.action == "tasks":
 		index = 0
 		cfg.dataset.tasks_list = args.tasks.split(",")

vall_e/models/__init__.py

@@ -15,6 +15,7 @@ def get_model(cfg):
 		d_model=cfg.dim,
 		n_heads=cfg.heads,
 		n_layers=cfg.layers,
+		config = cfg
 	)
 	model._cfg = cfg
 

vall_e/models/ar.py

@@ -22,8 +22,8 @@ class AR(Base):
 		return "ln"
 
 	@property
-	def arch_type(self) -> bool:
+	def arch_type(self) -> str:
-		if hasattr(self, "_cfg"):
+		if hasattr(self, "_cfg") and self._cfg:
 			return self._cfg.arch_type
 		return cfg.models.ar.arch_type
 
@@ -33,7 +33,7 @@ class AR(Base):
 
 	@property
 	def n_resp_levels(self) -> int:
-		if hasattr(self, "_cfg"):
+		if hasattr(self, "_cfg") and self._cfg:
 			return self._cfg.resp_levels
 		return cfg.models.ar.resp_levels
 
@@ -146,8 +146,8 @@ def example_usage():
 		tokenize("ˈ a ɪ   w ɪ l   nˌ ɑː t  ˈ æ s k   ɐ   sˈ ɛ k ə n d   tˈ a ɪ m").to(device),
 	]
 	proms_list = [
-		x8(torch.tensor([1, 2, 3], device=device)),
+		#x8(torch.tensor([1, 2, 3], device=device)),
-		#qnt.to(device),
+		qnt.to(device),
 	]
 	resps_list = [
 		qnt.to(device),
@@ -161,7 +161,7 @@ def example_usage():
 		'n_tokens': 1024,
 		'd_model': 1024,
 		'n_heads': 16,
-		'n_layers': 12,
+		'n_layers': 24,
 	}
 	model = AR(**kwargs).to(device)
 	engine = Engine(model=model, optimizer=torch.optim.AdamW(model.parameters(), lr=1e-4))

vall_e/models/nar.py

@@ -16,8 +16,8 @@ class NAR(Base):
 		return False
 
 	@property
-	def arch_type(self) -> bool:
+	def arch_type(self) -> str:
-		if hasattr(self, "_cfg"):
+		if hasattr(self, "_cfg") and self._cfg:
 			return self._cfg.arch_type
 		return cfg.models.nar.arch_type
 
@@ -31,7 +31,7 @@ class NAR(Base):
 
 	@property
 	def n_resp_levels(self) -> int:
-		if hasattr(self, "_cfg"):
+		if hasattr(self, "_cfg") and self._cfg:
 			return self._cfg.resp_levels
 		return cfg.models.nar.resp_levels
 

vall_e/utils/sampler.py

@@ -1,2 +1,29 @@
+from dataclasses import dataclass
+from typing import Any
+import random 
+
+@dataclass
 class Sampler():
-	...
+	def __init__( self, pool = [], keep_all = False ):
+		self.global_pool = pool if keep_all else None
+		self.global_indices = [ i for i in range(len(pool)) ]
+		self.reset()
+
+	def reset(self):
+		self.current_pool = [ i for i in self.global_indices ]
+
+	def sample(self, pool = None):
+		if pool is None:
+			pool = self.global_pool
+		# check if we need to reset 
+		index = random.choice( self.current_pool )
+		# remove from pool
+		self.current_pool.remove(index)
+		# reset if needed
+		if len(self.current_pool) == 0:
+			self.reset()
+		# map indices to our real values
+		return pool[index] if pool is not None else index
+
+	def __call__(self, *args, **kwargs):
+		return self.sample(*args, **kwargs)

vall_e/utils/trainer.py

@@ -311,6 +311,7 @@ def train(
 					print("Failed to set LR rate to:", rate, str(e))
 
 			if "export" in command:
+				train_dl.dataset.save_state_dict(cfg.relpath / "train_dataset.pt")
 				engines.save_checkpoint()
 				last_save_step = engines.global_step
 
@@ -333,6 +334,7 @@ def train(
 
 			if engines.global_step != last_save_step:
 				if engines.global_step % save_ckpt_every == 0 or command in saving_commands:
+					train_dl.dataset.save_state_dict(cfg.relpath / "train_dataset.pt")
 					engines.save_checkpoint()
 					last_save_step = engines.global_step